Keynotes

Keynote I: Automatic repair of advanced software systems: from AI-enabled to quantum systems

Development of advanced software systems, such as AI-enabled systems (e.g., automated driving systems) and quantum programs, is challenging; for example, AI-enabled systems may encounter, during operation, inputs that were not seen during training. For classical code, automated program repair (APR) has proven highly effective in addressing known failures by first identifying faults through fault localization and then using search-based approaches to generate repairs. The talk will provide an overview of a series of works that explore the adoption of automated repair for different types of advanced software systems, including classical DNNs, AI-enabled systems, and quantum programs. For DNNs, the talk will present three repair works: one that considers different types of misbehaviours, one that adjusts the repair target during the search, and one that considers orthogonal properties such as fairness. For AI-enabled systems, the talk will discuss approaches that account for the temporal nature of these systems during fault localization, as well as a study assessing the feasibility of open-loop repair. Finally, the talk will introduce a work investigating automated repair for quantum programs, where the fault localization phase is particularly challenging.

Invited Talk: AgentOps: Toward Reliable Operations for AI Agent Systems

AI agents are rapidly evolving from conversational assistants into autonomous systems that reason, invoke tools, interact with external environments, and execute complex tasks. While these capabilities enable increasingly sophisticated applications, they also introduce new reliability challenges. Failures in agent systems can emerge from multi-step reasoning, tool interactions, accumulated execution states, and dynamically evolving environments, making conventional monitoring and debugging techniques insufficient. This motivates AgentOps: a reliability-oriented paradigm for making agent systems observable, diagnosable, recoverable, and governable throughout their execution lifecycle.

This talk presents recent efforts toward building such an AgentOps framework through three complementary studies. The first investigates how failures in multi-agent systems can be systematically attributed to responsible components and interactions. The second explores structured recovery mechanisms that translate failure diagnosis into effective corrective actions. The third studies runtime intervention for stateful agents, enabling reliability guidance during ongoing executions rather than only after failures occur. Together, these studies form a progression from failure attribution to recovery to runtime reliability control. The talk concludes with open challenges in system-level coordination, fine-grained causal diagnosis, and adaptive closed-loop evolution, as well as the broader role of AgentOps in building dependable autonomous software systems.